drivers/infiniband/hw/bnxt_re/main.c

   1 /*
   2  * Broadcom NetXtreme-E RoCE driver.
   3  *
   4  * Copyright (c) 2016 - 2017, Broadcom. All rights reserved.  The term
   5  * Broadcom refers to Broadcom Limited and/or its subsidiaries.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * BSD license below:
  12  *
  13  * Redistribution and use in source and binary forms, with or without
  14  * modification, are permitted provided that the following conditions
  15  * are met:
  16  *
  17  * 1. Redistributions of source code must retain the above copyright
  18  *    notice, this list of conditions and the following disclaimer.
  19  * 2. Redistributions in binary form must reproduce the above copyright
  20  *    notice, this list of conditions and the following disclaimer in
  21  *    the documentation and/or other materials provided with the
  22  *    distribution.
  23  *
  24  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS''
  25  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
  26  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  27  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS
  28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
  30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
  31  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
  32  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
  33  * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN
  34  * IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  35  *
  36  * Description: Main component of the bnxt_re driver
  37  */
  38
  39 #include <linux/module.h>
  40 #include <linux/netdevice.h>
  41 #include <linux/ethtool.h>
  42 #include <linux/mutex.h>
  43 #include <linux/list.h>
  44 #include <linux/rculist.h>
  45 #include <linux/spinlock.h>
  46 #include <linux/pci.h>
  47 #include <net/dcbnl.h>
  48 #include <net/ipv6.h>
  49 #include <net/addrconf.h>
  50 #include <linux/if_ether.h>
  51
  52 #include <rdma/ib_verbs.h>
  53 #include <rdma/ib_user_verbs.h>
  54 #include <rdma/ib_umem.h>
  55 #include <rdma/ib_addr.h>
  56
  57 #include "bnxt_ulp.h"
  58 #include "roce_hsi.h"
  59 #include "qplib_res.h"
  60 #include "qplib_sp.h"
  61 #include "qplib_fp.h"
  62 #include "qplib_rcfw.h"
  63 #include "bnxt_re.h"
  64 #include "ib_verbs.h"
  65 #include <rdma/bnxt_re-abi.h>
  66 #include "bnxt.h"
  67 #include "hw_counters.h"
  68
  69 static char version[] =
  70                 BNXT_RE_DESC "\n";
  71
  72 MODULE_AUTHOR("Eddie Wai <eddie.wai@broadcom.com>");
  73 MODULE_DESCRIPTION(BNXT_RE_DESC " Driver");
  74 MODULE_LICENSE("Dual BSD/GPL");
  75
  76 /* globals */
  77 static struct list_head bnxt_re_dev_list = LIST_HEAD_INIT(bnxt_re_dev_list);
  78 /* Mutex to protect the list of bnxt_re devices added */
  79 static DEFINE_MUTEX(bnxt_re_dev_lock);
  80 static struct workqueue_struct *bnxt_re_wq;
  81 static void bnxt_re_ib_unreg(struct bnxt_re_dev *rdev);
  82
  83 static void bnxt_re_destroy_chip_ctx(struct bnxt_re_dev *rdev)
  84 {
  85         rdev->rcfw.res = NULL;
  86         rdev->qplib_res.cctx = NULL;
  87 }
  88
  89 static int bnxt_re_setup_chip_ctx(struct bnxt_re_dev *rdev)
  90 {
  91         struct bnxt_en_dev *en_dev;
  92         struct bnxt *bp;
  93
  94         en_dev = rdev->en_dev;
  95         bp = netdev_priv(en_dev->net);
  96
  97         rdev->chip_ctx.chip_num = bp->chip_num;
  98         /* rest members to follow eventually */
  99
 100         rdev->qplib_res.cctx = &rdev->chip_ctx;
 101         rdev->rcfw.res = &rdev->qplib_res;
 102
 103         return 0;
 104 }
 105
 106 /* SR-IOV helper functions */
 107
 108 static void bnxt_re_get_sriov_func_type(struct bnxt_re_dev *rdev)
 109 {
 110         struct bnxt *bp;
 111
 112         bp = netdev_priv(rdev->en_dev->net);
 113         if (BNXT_VF(bp))
 114                 rdev->is_virtfn = 1;
 115 }
 116
 117 /* Set the maximum number of each resource that the driver actually wants
 118  * to allocate. This may be up to the maximum number the firmware has
 119  * reserved for the function. The driver may choose to allocate fewer
 120  * resources than the firmware maximum.
 121  */
 122 static void bnxt_re_set_resource_limits(struct bnxt_re_dev *rdev)
 123 {
 124         u32 vf_qps = 0, vf_srqs = 0, vf_cqs = 0, vf_mrws = 0, vf_gids = 0;
 125         u32 i;
 126         u32 vf_pct;
 127         u32 num_vfs;
 128         struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
 129
 130         rdev->qplib_ctx.qpc_count = min_t(u32, BNXT_RE_MAX_QPC_COUNT,
 131                                           dev_attr->max_qp);
 132
 133         rdev->qplib_ctx.mrw_count = BNXT_RE_MAX_MRW_COUNT_256K;
 134         /* Use max_mr from fw since max_mrw does not get set */
 135         rdev->qplib_ctx.mrw_count = min_t(u32, rdev->qplib_ctx.mrw_count,
 136                                           dev_attr->max_mr);
 137         rdev->qplib_ctx.srqc_count = min_t(u32, BNXT_RE_MAX_SRQC_COUNT,
 138                                            dev_attr->max_srq);
 139         rdev->qplib_ctx.cq_count = min_t(u32, BNXT_RE_MAX_CQ_COUNT,
 140                                          dev_attr->max_cq);
 141
 142         for (i = 0; i < MAX_TQM_ALLOC_REQ; i++)
 143                 rdev->qplib_ctx.tqm_count[i] =
 144                 rdev->dev_attr.tqm_alloc_reqs[i];
 145
 146         if (rdev->num_vfs) {
 147                 /*
 148                  * Reserve a set of resources for the PF. Divide the remaining
 149                  * resources among the VFs
 150                  */
 151                 vf_pct = 100 - BNXT_RE_PCT_RSVD_FOR_PF;
 152                 num_vfs = 100 * rdev->num_vfs;
 153                 vf_qps = (rdev->qplib_ctx.qpc_count * vf_pct) / num_vfs;
 154                 vf_srqs = (rdev->qplib_ctx.srqc_count * vf_pct) / num_vfs;
 155                 vf_cqs = (rdev->qplib_ctx.cq_count * vf_pct) / num_vfs;
 156                 /*
 157                  * The driver allows many more MRs than other resources. If the
 158                  * firmware does also, then reserve a fixed amount for the PF
 159                  * and divide the rest among VFs. VFs may use many MRs for NFS
 160                  * mounts, ISER, NVME applications, etc. If the firmware
 161                  * severely restricts the number of MRs, then let PF have
 162                  * half and divide the rest among VFs, as for the other
 163                  * resource types.
 164                  */
 165                 if (rdev->qplib_ctx.mrw_count < BNXT_RE_MAX_MRW_COUNT_64K)
 166                         vf_mrws = rdev->qplib_ctx.mrw_count * vf_pct / num_vfs;
 167                 else
 168                         vf_mrws = (rdev->qplib_ctx.mrw_count -
 169                                    BNXT_RE_RESVD_MR_FOR_PF) / rdev->num_vfs;
 170                 vf_gids = BNXT_RE_MAX_GID_PER_VF;
 171         }
 172         rdev->qplib_ctx.vf_res.max_mrw_per_vf = vf_mrws;
 173         rdev->qplib_ctx.vf_res.max_gid_per_vf = vf_gids;
 174         rdev->qplib_ctx.vf_res.max_qp_per_vf = vf_qps;
 175         rdev->qplib_ctx.vf_res.max_srq_per_vf = vf_srqs;
 176         rdev->qplib_ctx.vf_res.max_cq_per_vf = vf_cqs;
 177 }
 178
 179 /* for handling bnxt_en callbacks later */
 180 static void bnxt_re_stop(void *p)
 181 {
 182 }
 183
 184 static void bnxt_re_start(void *p)
 185 {
 186 }
 187
 188 static void bnxt_re_sriov_config(void *p, int num_vfs)
 189 {
 190         struct bnxt_re_dev *rdev = p;
 191
 192         if (!rdev)
 193                 return;
 194
 195         rdev->num_vfs = num_vfs;
 196         bnxt_re_set_resource_limits(rdev);
 197         bnxt_qplib_set_func_resources(&rdev->qplib_res, &rdev->rcfw,
 198                                       &rdev->qplib_ctx);
 199 }
 200
 201 static void bnxt_re_shutdown(void *p)
 202 {
 203         struct bnxt_re_dev *rdev = p;
 204
 205         if (!rdev)
 206                 return;
 207
 208         bnxt_re_ib_unreg(rdev);
 209 }
 210
 211 static void bnxt_re_stop_irq(void *handle)
 212 {
 213         struct bnxt_re_dev *rdev = (struct bnxt_re_dev *)handle;
 214         struct bnxt_qplib_rcfw *rcfw = &rdev->rcfw;
 215         struct bnxt_qplib_nq *nq;
 216         int indx;
 217
 218         for (indx = BNXT_RE_NQ_IDX; indx < rdev->num_msix; indx++) {
 219                 nq = &rdev->nq[indx - 1];
 220                 bnxt_qplib_nq_stop_irq(nq, false);
 221         }
 222
 223         bnxt_qplib_rcfw_stop_irq(rcfw, false);
 224 }
 225
 226 static void bnxt_re_start_irq(void *handle, struct bnxt_msix_entry *ent)
 227 {
 228         struct bnxt_re_dev *rdev = (struct bnxt_re_dev *)handle;
 229         struct bnxt_msix_entry *msix_ent = rdev->msix_entries;
 230         struct bnxt_qplib_rcfw *rcfw = &rdev->rcfw;
 231         struct bnxt_qplib_nq *nq;
 232         int indx, rc;
 233
 234         if (!ent) {
 235                 /* Not setting the f/w timeout bit in rcfw.
 236                  * During the driver unload the first command
 237                  * to f/w will timeout and that will set the
 238                  * timeout bit.
 239                  */
 240                 dev_err(rdev_to_dev(rdev), "Failed to re-start IRQs\n");
 241                 return;
 242         }
 243
 244         /* Vectors may change after restart, so update with new vectors
 245          * in device sctructure.
 246          */
 247         for (indx = 0; indx < rdev->num_msix; indx++)
 248                 rdev->msix_entries[indx].vector = ent[indx].vector;
 249
 250         bnxt_qplib_rcfw_start_irq(rcfw, msix_ent[BNXT_RE_AEQ_IDX].vector,
 251                                   false);
 252         for (indx = BNXT_RE_NQ_IDX ; indx < rdev->num_msix; indx++) {
 253                 nq = &rdev->nq[indx - 1];
 254                 rc = bnxt_qplib_nq_start_irq(nq, indx - 1,
 255                                              msix_ent[indx].vector, false);
 256                 if (rc)
 257                         dev_warn(rdev_to_dev(rdev),
 258                                  "Failed to reinit NQ index %d\n", indx - 1);
 259         }
 260 }
 261
 262 static struct bnxt_ulp_ops bnxt_re_ulp_ops = {
 263         .ulp_async_notifier = NULL,
 264         .ulp_stop = bnxt_re_stop,
 265         .ulp_start = bnxt_re_start,
 266         .ulp_sriov_config = bnxt_re_sriov_config,
 267         .ulp_shutdown = bnxt_re_shutdown,
 268         .ulp_irq_stop = bnxt_re_stop_irq,
 269         .ulp_irq_restart = bnxt_re_start_irq
 270 };
 271
 272 /* RoCE -> Net driver */
 273
 274 /* Driver registration routines used to let the networking driver (bnxt_en)
 275  * to know that the RoCE driver is now installed
 276  */
 277 static int bnxt_re_unregister_netdev(struct bnxt_re_dev *rdev)
 278 {
 279         struct bnxt_en_dev *en_dev;
 280         int rc;
 281
 282         if (!rdev)
 283                 return -EINVAL;
 284
 285         en_dev = rdev->en_dev;
 286
 287         rc = en_dev->en_ops->bnxt_unregister_device(rdev->en_dev,
 288                                                     BNXT_ROCE_ULP);
 289         return rc;
 290 }
 291
 292 static int bnxt_re_register_netdev(struct bnxt_re_dev *rdev)
 293 {
 294         struct bnxt_en_dev *en_dev;
 295         int rc = 0;
 296
 297         if (!rdev)
 298                 return -EINVAL;
 299
 300         en_dev = rdev->en_dev;
 301
 302         rc = en_dev->en_ops->bnxt_register_device(en_dev, BNXT_ROCE_ULP,
 303                                                   &bnxt_re_ulp_ops, rdev);
 304         rdev->qplib_res.pdev = rdev->en_dev->pdev;
 305         return rc;
 306 }
 307
 308 static int bnxt_re_free_msix(struct bnxt_re_dev *rdev)
 309 {
 310         struct bnxt_en_dev *en_dev;
 311         int rc;
 312
 313         if (!rdev)
 314                 return -EINVAL;
 315
 316         en_dev = rdev->en_dev;
 317
 318
 319         rc = en_dev->en_ops->bnxt_free_msix(rdev->en_dev, BNXT_ROCE_ULP);
 320
 321         return rc;
 322 }
 323
 324 static int bnxt_re_request_msix(struct bnxt_re_dev *rdev)
 325 {
 326         int rc = 0, num_msix_want = BNXT_RE_MAX_MSIX, num_msix_got;
 327         struct bnxt_en_dev *en_dev;
 328
 329         if (!rdev)
 330                 return -EINVAL;
 331
 332         en_dev = rdev->en_dev;
 333
 334         num_msix_want = min_t(u32, BNXT_RE_MAX_MSIX, num_online_cpus());
 335
 336         num_msix_got = en_dev->en_ops->bnxt_request_msix(en_dev, BNXT_ROCE_ULP,
 337                                                          rdev->msix_entries,
 338                                                          num_msix_want);
 339         if (num_msix_got < BNXT_RE_MIN_MSIX) {
 340                 rc = -EINVAL;
 341                 goto done;
 342         }
 343         if (num_msix_got != num_msix_want) {
 344                 dev_warn(rdev_to_dev(rdev),
 345                          "Requested %d MSI-X vectors, got %d\n",
 346                          num_msix_want, num_msix_got);
 347         }
 348         rdev->num_msix = num_msix_got;
 349 done:
 350         return rc;
 351 }
 352
 353 static void bnxt_re_init_hwrm_hdr(struct bnxt_re_dev *rdev, struct input *hdr,
 354                                   u16 opcd, u16 crid, u16 trid)
 355 {
 356         hdr->req_type = cpu_to_le16(opcd);
 357         hdr->cmpl_ring = cpu_to_le16(crid);
 358         hdr->target_id = cpu_to_le16(trid);
 359 }
 360
 361 static void bnxt_re_fill_fw_msg(struct bnxt_fw_msg *fw_msg, void *msg,
 362                                 int msg_len, void *resp, int resp_max_len,
 363                                 int timeout)
 364 {
 365         fw_msg->msg = msg;
 366         fw_msg->msg_len = msg_len;
 367         fw_msg->resp = resp;
 368         fw_msg->resp_max_len = resp_max_len;
 369         fw_msg->timeout = timeout;
 370 }
 371
 372 static int bnxt_re_net_ring_free(struct bnxt_re_dev *rdev,
 373                                  u16 fw_ring_id, int type)
 374 {
 375         struct bnxt_en_dev *en_dev = rdev->en_dev;
 376         struct hwrm_ring_free_input req = {0};
 377         struct hwrm_ring_free_output resp;
 378         struct bnxt_fw_msg fw_msg;
 379         int rc = -EINVAL;
 380
 381         if (!en_dev)
 382                 return rc;
 383
 384         memset(&fw_msg, 0, sizeof(fw_msg));
 385
 386         bnxt_re_init_hwrm_hdr(rdev, (void *)&req, HWRM_RING_FREE, -1, -1);
 387         req.ring_type = type;
 388         req.ring_id = cpu_to_le16(fw_ring_id);
 389         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&resp,
 390                             sizeof(resp), DFLT_HWRM_CMD_TIMEOUT);
 391         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
 392         if (rc)
 393                 dev_err(rdev_to_dev(rdev),
 394                         "Failed to free HW ring:%d :%#x", req.ring_id, rc);
 395         return rc;
 396 }
 397
 398 static int bnxt_re_net_ring_alloc(struct bnxt_re_dev *rdev, dma_addr_t *dma_arr,
 399                                   int pages, int type, u32 ring_mask,
 400                                   u32 map_index, u16 *fw_ring_id)
 401 {
 402         struct bnxt_en_dev *en_dev = rdev->en_dev;
 403         struct hwrm_ring_alloc_input req = {0};
 404         struct hwrm_ring_alloc_output resp;
 405         struct bnxt_fw_msg fw_msg;
 406         int rc = -EINVAL;
 407
 408         if (!en_dev)
 409                 return rc;
 410
 411         memset(&fw_msg, 0, sizeof(fw_msg));
 412         bnxt_re_init_hwrm_hdr(rdev, (void *)&req, HWRM_RING_ALLOC, -1, -1);
 413         req.enables = 0;
 414         req.page_tbl_addr =  cpu_to_le64(dma_arr[0]);
 415         if (pages > 1) {
 416                 /* Page size is in log2 units */
 417                 req.page_size = BNXT_PAGE_SHIFT;
 418                 req.page_tbl_depth = 1;
 419         }
 420         req.fbo = 0;
 421         /* Association of ring index with doorbell index and MSIX number */
 422         req.logical_id = cpu_to_le16(map_index);
 423         req.length = cpu_to_le32(ring_mask + 1);
 424         req.ring_type = type;
 425         req.int_mode = RING_ALLOC_REQ_INT_MODE_MSIX;
 426         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&resp,
 427                             sizeof(resp), DFLT_HWRM_CMD_TIMEOUT);
 428         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
 429         if (!rc)
 430                 *fw_ring_id = le16_to_cpu(resp.ring_id);
 431
 432         return rc;
 433 }
 434
 435 static int bnxt_re_net_stats_ctx_free(struct bnxt_re_dev *rdev,
 436                                       u32 fw_stats_ctx_id)
 437 {
 438         struct bnxt_en_dev *en_dev = rdev->en_dev;
 439         struct hwrm_stat_ctx_free_input req = {0};
 440         struct bnxt_fw_msg fw_msg;
 441         int rc = -EINVAL;
 442
 443         if (!en_dev)
 444                 return rc;
 445
 446         memset(&fw_msg, 0, sizeof(fw_msg));
 447
 448         bnxt_re_init_hwrm_hdr(rdev, (void *)&req, HWRM_STAT_CTX_FREE, -1, -1);
 449         req.stat_ctx_id = cpu_to_le32(fw_stats_ctx_id);
 450         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&req,
 451                             sizeof(req), DFLT_HWRM_CMD_TIMEOUT);
 452         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
 453         if (rc)
 454                 dev_err(rdev_to_dev(rdev),
 455                         "Failed to free HW stats context %#x", rc);
 456
 457         return rc;
 458 }
 459
 460 static int bnxt_re_net_stats_ctx_alloc(struct bnxt_re_dev *rdev,
 461                                        dma_addr_t dma_map,
 462                                        u32 *fw_stats_ctx_id)
 463 {
 464         struct hwrm_stat_ctx_alloc_output resp = {0};
 465         struct hwrm_stat_ctx_alloc_input req = {0};
 466         struct bnxt_en_dev *en_dev = rdev->en_dev;
 467         struct bnxt_fw_msg fw_msg;
 468         int rc = -EINVAL;
 469
 470         *fw_stats_ctx_id = INVALID_STATS_CTX_ID;
 471
 472         if (!en_dev)
 473                 return rc;
 474
 475         memset(&fw_msg, 0, sizeof(fw_msg));
 476
 477         bnxt_re_init_hwrm_hdr(rdev, (void *)&req, HWRM_STAT_CTX_ALLOC, -1, -1);
 478         req.update_period_ms = cpu_to_le32(1000);
 479         req.stats_dma_addr = cpu_to_le64(dma_map);
 480         req.stat_ctx_flags = STAT_CTX_ALLOC_REQ_STAT_CTX_FLAGS_ROCE;
 481         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&resp,
 482                             sizeof(resp), DFLT_HWRM_CMD_TIMEOUT);
 483         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
 484         if (!rc)
 485                 *fw_stats_ctx_id = le32_to_cpu(resp.stat_ctx_id);
 486
 487         return rc;
 488 }
 489
 490 /* Device */
 491
 492 static bool is_bnxt_re_dev(struct net_device *netdev)
 493 {
 494         struct ethtool_drvinfo drvinfo;
 495
 496         if (netdev->ethtool_ops && netdev->ethtool_ops->get_drvinfo) {
 497                 memset(&drvinfo, 0, sizeof(drvinfo));
 498                 netdev->ethtool_ops->get_drvinfo(netdev, &drvinfo);
 499
 500                 if (strcmp(drvinfo.driver, "bnxt_en"))
 501                         return false;
 502                 return true;
 503         }
 504         return false;
 505 }
 506
 507 static struct bnxt_re_dev *bnxt_re_from_netdev(struct net_device *netdev)
 508 {
 509         struct bnxt_re_dev *rdev;
 510
 511         rcu_read_lock();
 512         list_for_each_entry_rcu(rdev, &bnxt_re_dev_list, list) {
 513                 if (rdev->netdev == netdev) {
 514                         rcu_read_unlock();
 515                         return rdev;
 516                 }
 517         }
 518         rcu_read_unlock();
 519         return NULL;
 520 }
 521
 522 static void bnxt_re_dev_unprobe(struct net_device *netdev,
 523                                 struct bnxt_en_dev *en_dev)
 524 {
 525         dev_put(netdev);
 526         module_put(en_dev->pdev->driver->driver.owner);
 527 }
 528
 529 static struct bnxt_en_dev *bnxt_re_dev_probe(struct net_device *netdev)
 530 {
 531         struct bnxt *bp = netdev_priv(netdev);
 532         struct bnxt_en_dev *en_dev;
 533         struct pci_dev *pdev;
 534
 535         /* Call bnxt_en's RoCE probe via indirect API */
 536         if (!bp->ulp_probe)
 537                 return ERR_PTR(-EINVAL);
 538
 539         en_dev = bp->ulp_probe(netdev);
 540         if (IS_ERR(en_dev))
 541                 return en_dev;
 542
 543         pdev = en_dev->pdev;
 544         if (!pdev)
 545                 return ERR_PTR(-EINVAL);
 546
 547         if (!(en_dev->flags & BNXT_EN_FLAG_ROCE_CAP)) {
 548                 dev_info(&pdev->dev,
 549                         "%s: probe error: RoCE is not supported on this device",
 550                         ROCE_DRV_MODULE_NAME);
 551                 return ERR_PTR(-ENODEV);
 552         }
 553
 554         /* Bump net device reference count */
 555         if (!try_module_get(pdev->driver->driver.owner))
 556                 return ERR_PTR(-ENODEV);
 557
 558         dev_hold(netdev);
 559
 560         return en_dev;
 561 }
 562
 563 static ssize_t hw_rev_show(struct device *device, struct device_attribute *attr,
 564                            char *buf)
 565 {
 566         struct bnxt_re_dev *rdev =
 567                 rdma_device_to_drv_device(device, struct bnxt_re_dev, ibdev);
 568
 569         return scnprintf(buf, PAGE_SIZE, "0x%x\n", rdev->en_dev->pdev->vendor);
 570 }
 571 static DEVICE_ATTR_RO(hw_rev);
 572
 573 static ssize_t hca_type_show(struct device *device,
 574                              struct device_attribute *attr, char *buf)
 575 {
 576         struct bnxt_re_dev *rdev =
 577                 rdma_device_to_drv_device(device, struct bnxt_re_dev, ibdev);
 578
 579         return scnprintf(buf, PAGE_SIZE, "%s\n", rdev->ibdev.node_desc);
 580 }
 581 static DEVICE_ATTR_RO(hca_type);
 582
 583 static struct attribute *bnxt_re_attributes[] = {
 584         &dev_attr_hw_rev.attr,
 585         &dev_attr_hca_type.attr,
 586         NULL
 587 };
 588
 589 static const struct attribute_group bnxt_re_dev_attr_group = {
 590         .attrs = bnxt_re_attributes,
 591 };
 592
 593 static void bnxt_re_unregister_ib(struct bnxt_re_dev *rdev)
 594 {
 595         ib_unregister_device(&rdev->ibdev);
 596 }
 597
 598 static const struct ib_device_ops bnxt_re_dev_ops = {
 599         .add_gid = bnxt_re_add_gid,
 600         .alloc_hw_stats = bnxt_re_ib_alloc_hw_stats,
 601         .alloc_mr = bnxt_re_alloc_mr,
 602         .alloc_pd = bnxt_re_alloc_pd,
 603         .alloc_ucontext = bnxt_re_alloc_ucontext,
 604         .create_ah = bnxt_re_create_ah,
 605         .create_cq = bnxt_re_create_cq,
 606         .create_qp = bnxt_re_create_qp,
 607         .create_srq = bnxt_re_create_srq,
 608         .dealloc_pd = bnxt_re_dealloc_pd,
 609         .dealloc_ucontext = bnxt_re_dealloc_ucontext,
 610         .del_gid = bnxt_re_del_gid,
 611         .dereg_mr = bnxt_re_dereg_mr,
 612         .destroy_ah = bnxt_re_destroy_ah,
 613         .destroy_cq = bnxt_re_destroy_cq,
 614         .destroy_qp = bnxt_re_destroy_qp,
 615         .destroy_srq = bnxt_re_destroy_srq,
 616         .get_dev_fw_str = bnxt_re_query_fw_str,
 617         .get_dma_mr = bnxt_re_get_dma_mr,
 618         .get_hw_stats = bnxt_re_ib_get_hw_stats,
 619         .get_link_layer = bnxt_re_get_link_layer,
 620         .get_port_immutable = bnxt_re_get_port_immutable,
 621         .map_mr_sg = bnxt_re_map_mr_sg,
 622         .mmap = bnxt_re_mmap,
 623         .modify_ah = bnxt_re_modify_ah,
 624         .modify_device = bnxt_re_modify_device,
 625         .modify_qp = bnxt_re_modify_qp,
 626         .modify_srq = bnxt_re_modify_srq,
 627         .poll_cq = bnxt_re_poll_cq,
 628         .post_recv = bnxt_re_post_recv,
 629         .post_send = bnxt_re_post_send,
 630         .post_srq_recv = bnxt_re_post_srq_recv,
 631         .query_ah = bnxt_re_query_ah,
 632         .query_device = bnxt_re_query_device,
 633         .query_pkey = bnxt_re_query_pkey,
 634         .query_port = bnxt_re_query_port,
 635         .query_qp = bnxt_re_query_qp,
 636         .query_srq = bnxt_re_query_srq,
 637         .reg_user_mr = bnxt_re_reg_user_mr,
 638         .req_notify_cq = bnxt_re_req_notify_cq,
 639         INIT_RDMA_OBJ_SIZE(ib_ah, bnxt_re_ah, ib_ah),
 640         INIT_RDMA_OBJ_SIZE(ib_pd, bnxt_re_pd, ib_pd),
 641         INIT_RDMA_OBJ_SIZE(ib_srq, bnxt_re_srq, ib_srq),
 642         INIT_RDMA_OBJ_SIZE(ib_ucontext, bnxt_re_ucontext, ib_uctx),
 643 };
 644
 645 static int bnxt_re_register_ib(struct bnxt_re_dev *rdev)
 646 {
 647         struct ib_device *ibdev = &rdev->ibdev;
 648         int ret;
 649
 650         /* ib device init */
 651         ibdev->owner = THIS_MODULE;
 652         ibdev->node_type = RDMA_NODE_IB_CA;
 653         strlcpy(ibdev->node_desc, BNXT_RE_DESC " HCA",
 654                 strlen(BNXT_RE_DESC) + 5);
 655         ibdev->phys_port_cnt = 1;
 656
 657         bnxt_qplib_get_guid(rdev->netdev->dev_addr, (u8 *)&ibdev->node_guid);
 658
 659         ibdev->num_comp_vectors = 1;
 660         ibdev->dev.parent = &rdev->en_dev->pdev->dev;
 661         ibdev->local_dma_lkey = BNXT_QPLIB_RSVD_LKEY;
 662
 663         /* User space */
 664         ibdev->uverbs_abi_ver = BNXT_RE_ABI_VERSION;
 665         ibdev->uverbs_cmd_mask =
 666                         (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
 667                         (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
 668                         (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
 669                         (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
 670                         (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
 671                         (1ull << IB_USER_VERBS_CMD_REG_MR)              |
 672                         (1ull << IB_USER_VERBS_CMD_REREG_MR)            |
 673                         (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
 674                         (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
 675                         (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
 676                         (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
 677                         (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
 678                         (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
 679                         (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
 680                         (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
 681                         (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
 682                         (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
 683                         (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
 684                         (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
 685                         (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
 686                         (1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
 687                         (1ull << IB_USER_VERBS_CMD_MODIFY_AH)           |
 688                         (1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
 689                         (1ull << IB_USER_VERBS_CMD_DESTROY_AH);
 690         /* POLL_CQ and REQ_NOTIFY_CQ is directly handled in libbnxt_re */
 691
 692
 693         rdma_set_device_sysfs_group(ibdev, &bnxt_re_dev_attr_group);
 694         ibdev->driver_id = RDMA_DRIVER_BNXT_RE;
 695         ib_set_device_ops(ibdev, &bnxt_re_dev_ops);
 696         ret = ib_device_set_netdev(&rdev->ibdev, rdev->netdev, 1);
 697         if (ret)
 698                 return ret;
 699
 700         return ib_register_device(ibdev, "bnxt_re%d");
 701 }
 702
 703 static void bnxt_re_dev_remove(struct bnxt_re_dev *rdev)
 704 {
 705         dev_put(rdev->netdev);
 706         rdev->netdev = NULL;
 707
 708         mutex_lock(&bnxt_re_dev_lock);
 709         list_del_rcu(&rdev->list);
 710         mutex_unlock(&bnxt_re_dev_lock);
 711
 712         synchronize_rcu();
 713
 714         ib_dealloc_device(&rdev->ibdev);
 715         /* rdev is gone */
 716 }
 717
 718 static struct bnxt_re_dev *bnxt_re_dev_add(struct net_device *netdev,
 719                                            struct bnxt_en_dev *en_dev)
 720 {
 721         struct bnxt_re_dev *rdev;
 722
 723         /* Allocate bnxt_re_dev instance here */
 724         rdev = ib_alloc_device(bnxt_re_dev, ibdev);
 725         if (!rdev) {
 726                 dev_err(NULL, "%s: bnxt_re_dev allocation failure!",
 727                         ROCE_DRV_MODULE_NAME);
 728                 return NULL;
 729         }
 730         /* Default values */
 731         rdev->netdev = netdev;
 732         dev_hold(rdev->netdev);
 733         rdev->en_dev = en_dev;
 734         rdev->id = rdev->en_dev->pdev->devfn;
 735         INIT_LIST_HEAD(&rdev->qp_list);
 736         mutex_init(&rdev->qp_lock);
 737         atomic_set(&rdev->qp_count, 0);
 738         atomic_set(&rdev->cq_count, 0);
 739         atomic_set(&rdev->srq_count, 0);
 740         atomic_set(&rdev->mr_count, 0);
 741         atomic_set(&rdev->mw_count, 0);
 742         rdev->cosq[0] = 0xFFFF;
 743         rdev->cosq[1] = 0xFFFF;
 744
 745         mutex_lock(&bnxt_re_dev_lock);
 746         list_add_tail_rcu(&rdev->list, &bnxt_re_dev_list);
 747         mutex_unlock(&bnxt_re_dev_lock);
 748         return rdev;
 749 }
 750
 751 static int bnxt_re_handle_unaffi_async_event(struct creq_func_event
 752                                              *unaffi_async)
 753 {
 754         switch (unaffi_async->event) {
 755         case CREQ_FUNC_EVENT_EVENT_TX_WQE_ERROR:
 756                 break;
 757         case CREQ_FUNC_EVENT_EVENT_TX_DATA_ERROR:
 758                 break;
 759         case CREQ_FUNC_EVENT_EVENT_RX_WQE_ERROR:
 760                 break;
 761         case CREQ_FUNC_EVENT_EVENT_RX_DATA_ERROR:
 762                 break;
 763         case CREQ_FUNC_EVENT_EVENT_CQ_ERROR:
 764                 break;
 765         case CREQ_FUNC_EVENT_EVENT_TQM_ERROR:
 766                 break;
 767         case CREQ_FUNC_EVENT_EVENT_CFCQ_ERROR:
 768                 break;
 769         case CREQ_FUNC_EVENT_EVENT_CFCS_ERROR:
 770                 break;
 771         case CREQ_FUNC_EVENT_EVENT_CFCC_ERROR:
 772                 break;
 773         case CREQ_FUNC_EVENT_EVENT_CFCM_ERROR:
 774                 break;
 775         case CREQ_FUNC_EVENT_EVENT_TIM_ERROR:
 776                 break;
 777         default:
 778                 return -EINVAL;
 779         }
 780         return 0;
 781 }
 782
 783 static int bnxt_re_handle_qp_async_event(struct creq_qp_event *qp_event,
 784                                          struct bnxt_re_qp *qp)
 785 {
 786         struct ib_event event;
 787         unsigned int flags;
 788
 789         if (qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR) {
 790                 flags = bnxt_re_lock_cqs(qp);
 791                 bnxt_qplib_add_flush_qp(&qp->qplib_qp);
 792                 bnxt_re_unlock_cqs(qp, flags);
 793         }
 794
 795         memset(&event, 0, sizeof(event));
 796         if (qp->qplib_qp.srq) {
 797                 event.device = &qp->rdev->ibdev;
 798                 event.element.qp = &qp->ib_qp;
 799                 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
 800         }
 801
 802         if (event.device && qp->ib_qp.event_handler)
 803                 qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
 804
 805         return 0;
 806 }
 807
 808 static int bnxt_re_handle_affi_async_event(struct creq_qp_event *affi_async,
 809                                            void *obj)
 810 {
 811         int rc = 0;
 812         u8 event;
 813
 814         if (!obj)
 815                 return rc; /* QP was already dead, still return success */
 816
 817         event = affi_async->event;
 818         if (event == CREQ_QP_EVENT_EVENT_QP_ERROR_NOTIFICATION) {
 819                 struct bnxt_qplib_qp *lib_qp = obj;
 820                 struct bnxt_re_qp *qp = container_of(lib_qp, struct bnxt_re_qp,
 821                                                      qplib_qp);
 822                 rc = bnxt_re_handle_qp_async_event(affi_async, qp);
 823         }
 824         return rc;
 825 }
 826
 827 static int bnxt_re_aeq_handler(struct bnxt_qplib_rcfw *rcfw,
 828                                void *aeqe, void *obj)
 829 {
 830         struct creq_qp_event *affi_async;
 831         struct creq_func_event *unaffi_async;
 832         u8 type;
 833         int rc;
 834
 835         type = ((struct creq_base *)aeqe)->type;
 836         if (type == CREQ_BASE_TYPE_FUNC_EVENT) {
 837                 unaffi_async = aeqe;
 838                 rc = bnxt_re_handle_unaffi_async_event(unaffi_async);
 839         } else {
 840                 affi_async = aeqe;
 841                 rc = bnxt_re_handle_affi_async_event(affi_async, obj);
 842         }
 843
 844         return rc;
 845 }
 846
 847 static int bnxt_re_srqn_handler(struct bnxt_qplib_nq *nq,
 848                                 struct bnxt_qplib_srq *handle, u8 event)
 849 {
 850         struct bnxt_re_srq *srq = container_of(handle, struct bnxt_re_srq,
 851                                                qplib_srq);
 852         struct ib_event ib_event;
 853         int rc = 0;
 854
 855         if (!srq) {
 856                 dev_err(NULL, "%s: SRQ is NULL, SRQN not handled",
 857                         ROCE_DRV_MODULE_NAME);
 858                 rc = -EINVAL;
 859                 goto done;
 860         }
 861         ib_event.device = &srq->rdev->ibdev;
 862         ib_event.element.srq = &srq->ib_srq;
 863         if (event == NQ_SRQ_EVENT_EVENT_SRQ_THRESHOLD_EVENT)
 864                 ib_event.event = IB_EVENT_SRQ_LIMIT_REACHED;
 865         else
 866                 ib_event.event = IB_EVENT_SRQ_ERR;
 867
 868         if (srq->ib_srq.event_handler) {
 869                 /* Lock event_handler? */
 870                 (*srq->ib_srq.event_handler)(&ib_event,
 871                                              srq->ib_srq.srq_context);
 872         }
 873 done:
 874         return rc;
 875 }
 876
 877 static int bnxt_re_cqn_handler(struct bnxt_qplib_nq *nq,
 878                                struct bnxt_qplib_cq *handle)
 879 {
 880         struct bnxt_re_cq *cq = container_of(handle, struct bnxt_re_cq,
 881                                              qplib_cq);
 882
 883         if (!cq) {
 884                 dev_err(NULL, "%s: CQ is NULL, CQN not handled",
 885                         ROCE_DRV_MODULE_NAME);
 886                 return -EINVAL;
 887         }
 888         if (cq->ib_cq.comp_handler) {
 889                 /* Lock comp_handler? */
 890                 (*cq->ib_cq.comp_handler)(&cq->ib_cq, cq->ib_cq.cq_context);
 891         }
 892
 893         return 0;
 894 }
 895
 896 static u32 bnxt_re_get_nqdb_offset(struct bnxt_re_dev *rdev, u16 indx)
 897 {
 898         return bnxt_qplib_is_chip_gen_p5(&rdev->chip_ctx) ?
 899                                 0x10000 : rdev->msix_entries[indx].db_offset;
 900 }
 901
 902 static void bnxt_re_cleanup_res(struct bnxt_re_dev *rdev)
 903 {
 904         int i;
 905
 906         for (i = 1; i < rdev->num_msix; i++)
 907                 bnxt_qplib_disable_nq(&rdev->nq[i - 1]);
 908
 909         if (rdev->qplib_res.rcfw)
 910                 bnxt_qplib_cleanup_res(&rdev->qplib_res);
 911 }
 912
 913 static int bnxt_re_init_res(struct bnxt_re_dev *rdev)
 914 {
 915         int num_vec_enabled = 0;
 916         int rc = 0, i;
 917         u32 db_offt;
 918
 919         bnxt_qplib_init_res(&rdev->qplib_res);
 920
 921         for (i = 1; i < rdev->num_msix ; i++) {
 922                 db_offt = bnxt_re_get_nqdb_offset(rdev, i);
 923                 rc = bnxt_qplib_enable_nq(rdev->en_dev->pdev, &rdev->nq[i - 1],
 924                                           i - 1, rdev->msix_entries[i].vector,
 925                                           db_offt, &bnxt_re_cqn_handler,
 926                                           &bnxt_re_srqn_handler);
 927                 if (rc) {
 928                         dev_err(rdev_to_dev(rdev),
 929                                 "Failed to enable NQ with rc = 0x%x", rc);
 930                         goto fail;
 931                 }
 932                 num_vec_enabled++;
 933         }
 934         return 0;
 935 fail:
 936         for (i = num_vec_enabled; i >= 0; i--)
 937                 bnxt_qplib_disable_nq(&rdev->nq[i]);
 938         return rc;
 939 }
 940
 941 static void bnxt_re_free_nq_res(struct bnxt_re_dev *rdev)
 942 {
 943         u8 type;
 944         int i;
 945
 946         for (i = 0; i < rdev->num_msix - 1; i++) {
 947                 type = bnxt_qplib_get_ring_type(&rdev->chip_ctx);
 948                 bnxt_re_net_ring_free(rdev, rdev->nq[i].ring_id, type);
 949                 rdev->nq[i].res = NULL;
 950                 bnxt_qplib_free_nq(&rdev->nq[i]);
 951         }
 952 }
 953
 954 static void bnxt_re_free_res(struct bnxt_re_dev *rdev)
 955 {
 956         bnxt_re_free_nq_res(rdev);
 957
 958         if (rdev->qplib_res.dpi_tbl.max) {
 959                 bnxt_qplib_dealloc_dpi(&rdev->qplib_res,
 960                                        &rdev->qplib_res.dpi_tbl,
 961                                        &rdev->dpi_privileged);
 962         }
 963         if (rdev->qplib_res.rcfw) {
 964                 bnxt_qplib_free_res(&rdev->qplib_res);
 965                 rdev->qplib_res.rcfw = NULL;
 966         }
 967 }
 968
 969 static int bnxt_re_alloc_res(struct bnxt_re_dev *rdev)
 970 {
 971         int num_vec_created = 0;
 972         dma_addr_t *pg_map;
 973         int rc = 0, i;
 974         int pages;
 975         u8 type;
 976
 977         /* Configure and allocate resources for qplib */
 978         rdev->qplib_res.rcfw = &rdev->rcfw;
 979         rc = bnxt_qplib_get_dev_attr(&rdev->rcfw, &rdev->dev_attr,
 980                                      rdev->is_virtfn);
 981         if (rc)
 982                 goto fail;
 983
 984         rc = bnxt_qplib_alloc_res(&rdev->qplib_res, rdev->en_dev->pdev,
 985                                   rdev->netdev, &rdev->dev_attr);
 986         if (rc)
 987                 goto fail;
 988
 989         rc = bnxt_qplib_alloc_dpi(&rdev->qplib_res.dpi_tbl,
 990                                   &rdev->dpi_privileged,
 991                                   rdev);
 992         if (rc)
 993                 goto dealloc_res;
 994
 995         for (i = 0; i < rdev->num_msix - 1; i++) {
 996                 rdev->nq[i].res = &rdev->qplib_res;
 997                 rdev->nq[i].hwq.max_elements = BNXT_RE_MAX_CQ_COUNT +
 998                         BNXT_RE_MAX_SRQC_COUNT + 2;
 999                 rc = bnxt_qplib_alloc_nq(rdev->en_dev->pdev, &rdev->nq[i]);
1000                 if (rc) {
1001                         dev_err(rdev_to_dev(rdev), "Alloc Failed NQ%d rc:%#x",
1002                                 i, rc);
1003                         goto free_nq;
1004                 }
1005                 type = bnxt_qplib_get_ring_type(&rdev->chip_ctx);
1006                 pg_map = rdev->nq[i].hwq.pbl[PBL_LVL_0].pg_map_arr;
1007                 pages = rdev->nq[i].hwq.pbl[rdev->nq[i].hwq.level].pg_count;
1008                 rc = bnxt_re_net_ring_alloc(rdev, pg_map, pages, type,
1009                                             BNXT_QPLIB_NQE_MAX_CNT - 1,
1010                                             rdev->msix_entries[i + 1].ring_idx,
1011                                             &rdev->nq[i].ring_id);
1012                 if (rc) {
1013                         dev_err(rdev_to_dev(rdev),
1014                                 "Failed to allocate NQ fw id with rc = 0x%x",
1015                                 rc);
1016                         bnxt_qplib_free_nq(&rdev->nq[i]);
1017                         goto free_nq;
1018                 }
1019                 num_vec_created++;
1020         }
1021         return 0;
1022 free_nq:
1023         for (i = num_vec_created; i >= 0; i--) {
1024                 type = bnxt_qplib_get_ring_type(&rdev->chip_ctx);
1025                 bnxt_re_net_ring_free(rdev, rdev->nq[i].ring_id, type);
1026                 bnxt_qplib_free_nq(&rdev->nq[i]);
1027         }
1028         bnxt_qplib_dealloc_dpi(&rdev->qplib_res,
1029                                &rdev->qplib_res.dpi_tbl,
1030                                &rdev->dpi_privileged);
1031 dealloc_res:
1032         bnxt_qplib_free_res(&rdev->qplib_res);
1033
1034 fail:
1035         rdev->qplib_res.rcfw = NULL;
1036         return rc;
1037 }
1038
1039 static void bnxt_re_dispatch_event(struct ib_device *ibdev, struct ib_qp *qp,
1040                                    u8 port_num, enum ib_event_type event)
1041 {
1042         struct ib_event ib_event;
1043
1044         ib_event.device = ibdev;
1045         if (qp) {
1046                 ib_event.element.qp = qp;
1047                 ib_event.event = event;
1048                 if (qp->event_handler)
1049                         qp->event_handler(&ib_event, qp->qp_context);
1050
1051         } else {
1052                 ib_event.element.port_num = port_num;
1053                 ib_event.event = event;
1054                 ib_dispatch_event(&ib_event);
1055         }
1056 }
1057
1058 #define HWRM_QUEUE_PRI2COS_QCFG_INPUT_FLAGS_IVLAN      0x02
1059 static int bnxt_re_query_hwrm_pri2cos(struct bnxt_re_dev *rdev, u8 dir,
1060                                       u64 *cid_map)
1061 {
1062         struct hwrm_queue_pri2cos_qcfg_input req = {0};
1063         struct bnxt *bp = netdev_priv(rdev->netdev);
1064         struct hwrm_queue_pri2cos_qcfg_output resp;
1065         struct bnxt_en_dev *en_dev = rdev->en_dev;
1066         struct bnxt_fw_msg fw_msg;
1067         u32 flags = 0;
1068         u8 *qcfgmap, *tmp_map;
1069         int rc = 0, i;
1070
1071         if (!cid_map)
1072                 return -EINVAL;
1073
1074         memset(&fw_msg, 0, sizeof(fw_msg));
1075         bnxt_re_init_hwrm_hdr(rdev, (void *)&req,
1076                               HWRM_QUEUE_PRI2COS_QCFG, -1, -1);
1077         flags |= (dir & 0x01);
1078         flags |= HWRM_QUEUE_PRI2COS_QCFG_INPUT_FLAGS_IVLAN;
1079         req.flags = cpu_to_le32(flags);
1080         req.port_id = bp->pf.port_id;
1081
1082         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&resp,
1083                             sizeof(resp), DFLT_HWRM_CMD_TIMEOUT);
1084         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
1085         if (rc)
1086                 return rc;
1087
1088         if (resp.queue_cfg_info) {
1089                 dev_warn(rdev_to_dev(rdev),
1090                          "Asymmetric cos queue configuration detected");
1091                 dev_warn(rdev_to_dev(rdev),
1092                          " on device, QoS may not be fully functional\n");
1093         }
1094         qcfgmap = &resp.pri0_cos_queue_id;
1095         tmp_map = (u8 *)cid_map;
1096         for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++)
1097                 tmp_map[i] = qcfgmap[i];
1098
1099         return rc;
1100 }
1101
1102 static bool bnxt_re_is_qp1_or_shadow_qp(struct bnxt_re_dev *rdev,
1103                                         struct bnxt_re_qp *qp)
1104 {
1105         return (qp->ib_qp.qp_type == IB_QPT_GSI) || (qp == rdev->qp1_sqp);
1106 }
1107
1108 static void bnxt_re_dev_stop(struct bnxt_re_dev *rdev)
1109 {
1110         int mask = IB_QP_STATE;
1111         struct ib_qp_attr qp_attr;
1112         struct bnxt_re_qp *qp;
1113
1114         qp_attr.qp_state = IB_QPS_ERR;
1115         mutex_lock(&rdev->qp_lock);
1116         list_for_each_entry(qp, &rdev->qp_list, list) {
1117                 /* Modify the state of all QPs except QP1/Shadow QP */
1118                 if (!bnxt_re_is_qp1_or_shadow_qp(rdev, qp)) {
1119                         if (qp->qplib_qp.state !=
1120                             CMDQ_MODIFY_QP_NEW_STATE_RESET &&
1121                             qp->qplib_qp.state !=
1122                             CMDQ_MODIFY_QP_NEW_STATE_ERR) {
1123                                 bnxt_re_dispatch_event(&rdev->ibdev, &qp->ib_qp,
1124                                                        1, IB_EVENT_QP_FATAL);
1125                                 bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, mask,
1126                                                   NULL);
1127                         }
1128                 }
1129         }
1130         mutex_unlock(&rdev->qp_lock);
1131 }
1132
1133 static int bnxt_re_update_gid(struct bnxt_re_dev *rdev)
1134 {
1135         struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
1136         struct bnxt_qplib_gid gid;
1137         u16 gid_idx, index;
1138         int rc = 0;
1139
1140         if (!test_bit(BNXT_RE_FLAG_IBDEV_REGISTERED, &rdev->flags))
1141                 return 0;
1142
1143         if (!sgid_tbl) {
1144                 dev_err(rdev_to_dev(rdev), "QPLIB: SGID table not allocated");
1145                 return -EINVAL;
1146         }
1147
1148         for (index = 0; index < sgid_tbl->active; index++) {
1149                 gid_idx = sgid_tbl->hw_id[index];
1150
1151                 if (!memcmp(&sgid_tbl->tbl[index], &bnxt_qplib_gid_zero,
1152                             sizeof(bnxt_qplib_gid_zero)))
1153                         continue;
1154                 /* need to modify the VLAN enable setting of non VLAN GID only
1155                  * as setting is done for VLAN GID while adding GID
1156                  */
1157                 if (sgid_tbl->vlan[index])
1158                         continue;
1159
1160                 memcpy(&gid, &sgid_tbl->tbl[index], sizeof(gid));
1161
1162                 rc = bnxt_qplib_update_sgid(sgid_tbl, &gid, gid_idx,
1163                                             rdev->qplib_res.netdev->dev_addr);
1164         }
1165
1166         return rc;
1167 }
1168
1169 static u32 bnxt_re_get_priority_mask(struct bnxt_re_dev *rdev)
1170 {
1171         u32 prio_map = 0, tmp_map = 0;
1172         struct net_device *netdev;
1173         struct dcb_app app;
1174
1175         netdev = rdev->netdev;
1176
1177         memset(&app, 0, sizeof(app));
1178         app.selector = IEEE_8021QAZ_APP_SEL_ETHERTYPE;
1179         app.protocol = ETH_P_IBOE;
1180         tmp_map = dcb_ieee_getapp_mask(netdev, &app);
1181         prio_map = tmp_map;
1182
1183         app.selector = IEEE_8021QAZ_APP_SEL_DGRAM;
1184         app.protocol = ROCE_V2_UDP_DPORT;
1185         tmp_map = dcb_ieee_getapp_mask(netdev, &app);
1186         prio_map |= tmp_map;
1187
1188         return prio_map;
1189 }
1190
1191 static void bnxt_re_parse_cid_map(u8 prio_map, u8 *cid_map, u16 *cosq)
1192 {
1193         u16 prio;
1194         u8 id;
1195
1196         for (prio = 0, id = 0; prio < 8; prio++) {
1197                 if (prio_map & (1 << prio)) {
1198                         cosq[id] = cid_map[prio];
1199                         id++;
1200                         if (id == 2) /* Max 2 tcs supported */
1201                                 break;
1202                 }
1203         }
1204 }
1205
1206 static int bnxt_re_setup_qos(struct bnxt_re_dev *rdev)
1207 {
1208         u8 prio_map = 0;
1209         u64 cid_map;
1210         int rc;
1211
1212         /* Get priority for roce */
1213         prio_map = bnxt_re_get_priority_mask(rdev);
1214
1215         if (prio_map == rdev->cur_prio_map)
1216                 return 0;
1217         rdev->cur_prio_map = prio_map;
1218         /* Get cosq id for this priority */
1219         rc = bnxt_re_query_hwrm_pri2cos(rdev, 0, &cid_map);
1220         if (rc) {
1221                 dev_warn(rdev_to_dev(rdev), "no cos for p_mask %x\n", prio_map);
1222                 return rc;
1223         }
1224         /* Parse CoS IDs for app priority */
1225         bnxt_re_parse_cid_map(prio_map, (u8 *)&cid_map, rdev->cosq);
1226
1227         /* Config BONO. */
1228         rc = bnxt_qplib_map_tc2cos(&rdev->qplib_res, rdev->cosq);
1229         if (rc) {
1230                 dev_warn(rdev_to_dev(rdev), "no tc for cos{%x, %x}\n",
1231                          rdev->cosq[0], rdev->cosq[1]);
1232                 return rc;
1233         }
1234
1235         /* Actual priorities are not programmed as they are already
1236          * done by L2 driver; just enable or disable priority vlan tagging
1237          */
1238         if ((prio_map == 0 && rdev->qplib_res.prio) ||
1239             (prio_map != 0 && !rdev->qplib_res.prio)) {
1240                 rdev->qplib_res.prio = prio_map ? true : false;
1241
1242                 bnxt_re_update_gid(rdev);
1243         }
1244
1245         return 0;
1246 }
1247
1248 static void bnxt_re_query_hwrm_intf_version(struct bnxt_re_dev *rdev)
1249 {
1250         struct bnxt_en_dev *en_dev = rdev->en_dev;
1251         struct hwrm_ver_get_output resp = {0};
1252         struct hwrm_ver_get_input req = {0};
1253         struct bnxt_fw_msg fw_msg;
1254         int rc = 0;
1255
1256         memset(&fw_msg, 0, sizeof(fw_msg));
1257         bnxt_re_init_hwrm_hdr(rdev, (void *)&req,
1258                               HWRM_VER_GET, -1, -1);
1259         req.hwrm_intf_maj = HWRM_VERSION_MAJOR;
1260         req.hwrm_intf_min = HWRM_VERSION_MINOR;
1261         req.hwrm_intf_upd = HWRM_VERSION_UPDATE;
1262         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&resp,
1263                             sizeof(resp), DFLT_HWRM_CMD_TIMEOUT);
1264         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
1265         if (rc) {
1266                 dev_err(rdev_to_dev(rdev),
1267                         "Failed to query HW version, rc = 0x%x", rc);
1268                 return;
1269         }
1270         rdev->qplib_ctx.hwrm_intf_ver =
1271                 (u64)resp.hwrm_intf_major << 48 |
1272                 (u64)resp.hwrm_intf_minor << 32 |
1273                 (u64)resp.hwrm_intf_build << 16 |
1274                 resp.hwrm_intf_patch;
1275 }
1276
1277 static void bnxt_re_ib_unreg(struct bnxt_re_dev *rdev)
1278 {
1279         u8 type;
1280         int rc;
1281
1282         if (test_and_clear_bit(BNXT_RE_FLAG_IBDEV_REGISTERED, &rdev->flags)) {
1283                 /* Cleanup ib dev */
1284                 bnxt_re_unregister_ib(rdev);
1285         }
1286         if (test_and_clear_bit(BNXT_RE_FLAG_QOS_WORK_REG, &rdev->flags))
1287                 cancel_delayed_work_sync(&rdev->worker);
1288
1289         if (test_and_clear_bit(BNXT_RE_FLAG_RESOURCES_INITIALIZED,
1290                                &rdev->flags))
1291                 bnxt_re_cleanup_res(rdev);
1292         if (test_and_clear_bit(BNXT_RE_FLAG_RESOURCES_ALLOCATED, &rdev->flags))
1293                 bnxt_re_free_res(rdev);
1294
1295         if (test_and_clear_bit(BNXT_RE_FLAG_RCFW_CHANNEL_EN, &rdev->flags)) {
1296                 rc = bnxt_qplib_deinit_rcfw(&rdev->rcfw);
1297                 if (rc)
1298                         dev_warn(rdev_to_dev(rdev),
1299                                  "Failed to deinitialize RCFW: %#x", rc);
1300                 bnxt_re_net_stats_ctx_free(rdev, rdev->qplib_ctx.stats.fw_id);
1301                 bnxt_qplib_free_ctx(rdev->en_dev->pdev, &rdev->qplib_ctx);
1302                 bnxt_qplib_disable_rcfw_channel(&rdev->rcfw);
1303                 type = bnxt_qplib_get_ring_type(&rdev->chip_ctx);
1304                 bnxt_re_net_ring_free(rdev, rdev->rcfw.creq_ring_id, type);
1305                 bnxt_qplib_free_rcfw_channel(&rdev->rcfw);
1306         }
1307         if (test_and_clear_bit(BNXT_RE_FLAG_GOT_MSIX, &rdev->flags)) {
1308                 rc = bnxt_re_free_msix(rdev);
1309                 if (rc)
1310                         dev_warn(rdev_to_dev(rdev),
1311                                  "Failed to free MSI-X vectors: %#x", rc);
1312         }
1313
1314         bnxt_re_destroy_chip_ctx(rdev);
1315         if (test_and_clear_bit(BNXT_RE_FLAG_NETDEV_REGISTERED, &rdev->flags)) {
1316                 rc = bnxt_re_unregister_netdev(rdev);
1317                 if (rc)
1318                         dev_warn(rdev_to_dev(rdev),
1319                                  "Failed to unregister with netdev: %#x", rc);
1320         }
1321 }
1322
1323 /* worker thread for polling periodic events. Now used for QoS programming*/
1324 static void bnxt_re_worker(struct work_struct *work)
1325 {
1326         struct bnxt_re_dev *rdev = container_of(work, struct bnxt_re_dev,
1327                                                 worker.work);
1328
1329         bnxt_re_setup_qos(rdev);
1330         schedule_delayed_work(&rdev->worker, msecs_to_jiffies(30000));
1331 }
1332
1333 static int bnxt_re_ib_reg(struct bnxt_re_dev *rdev)
1334 {
1335         dma_addr_t *pg_map;
1336         u32 db_offt, ridx;
1337         int pages, vid;
1338         bool locked;
1339         u8 type;
1340         int rc;
1341
1342         /* Acquire rtnl lock through out this function */
1343         rtnl_lock();
1344         locked = true;
1345
1346         /* Registered a new RoCE device instance to netdev */
1347         rc = bnxt_re_register_netdev(rdev);
1348         if (rc) {
1349                 rtnl_unlock();
1350                 pr_err("Failed to register with netedev: %#x\n", rc);
1351                 return -EINVAL;
1352         }
1353         set_bit(BNXT_RE_FLAG_NETDEV_REGISTERED, &rdev->flags);
1354
1355         rc = bnxt_re_setup_chip_ctx(rdev);
1356         if (rc) {
1357                 dev_err(rdev_to_dev(rdev), "Failed to get chip context\n");
1358                 return -EINVAL;
1359         }
1360
1361         /* Check whether VF or PF */
1362         bnxt_re_get_sriov_func_type(rdev);
1363
1364         rc = bnxt_re_request_msix(rdev);
1365         if (rc) {
1366                 pr_err("Failed to get MSI-X vectors: %#x\n", rc);
1367                 rc = -EINVAL;
1368                 goto fail;
1369         }
1370         set_bit(BNXT_RE_FLAG_GOT_MSIX, &rdev->flags);
1371
1372         bnxt_re_query_hwrm_intf_version(rdev);
1373
1374         /* Establish RCFW Communication Channel to initialize the context
1375          * memory for the function and all child VFs
1376          */
1377         rc = bnxt_qplib_alloc_rcfw_channel(rdev->en_dev->pdev, &rdev->rcfw,
1378                                            &rdev->qplib_ctx,
1379                                            BNXT_RE_MAX_QPC_COUNT);
1380         if (rc) {
1381                 pr_err("Failed to allocate RCFW Channel: %#x\n", rc);
1382                 goto fail;
1383         }
1384         type = bnxt_qplib_get_ring_type(&rdev->chip_ctx);
1385         pg_map = rdev->rcfw.creq.pbl[PBL_LVL_0].pg_map_arr;
1386         pages = rdev->rcfw.creq.pbl[rdev->rcfw.creq.level].pg_count;
1387         ridx = rdev->msix_entries[BNXT_RE_AEQ_IDX].ring_idx;
1388         rc = bnxt_re_net_ring_alloc(rdev, pg_map, pages, type,
1389                                     BNXT_QPLIB_CREQE_MAX_CNT - 1,
1390                                     ridx, &rdev->rcfw.creq_ring_id);
1391         if (rc) {
1392                 pr_err("Failed to allocate CREQ: %#x\n", rc);
1393                 goto free_rcfw;
1394         }
1395         db_offt = bnxt_re_get_nqdb_offset(rdev, BNXT_RE_AEQ_IDX);
1396         vid = rdev->msix_entries[BNXT_RE_AEQ_IDX].vector;
1397         rc = bnxt_qplib_enable_rcfw_channel(rdev->en_dev->pdev, &rdev->rcfw,
1398                                             vid, db_offt, rdev->is_virtfn,
1399                                             &bnxt_re_aeq_handler);
1400         if (rc) {
1401                 pr_err("Failed to enable RCFW channel: %#x\n", rc);
1402                 goto free_ring;
1403         }
1404
1405         rc = bnxt_qplib_get_dev_attr(&rdev->rcfw, &rdev->dev_attr,
1406                                      rdev->is_virtfn);
1407         if (rc)
1408                 goto disable_rcfw;
1409         if (!rdev->is_virtfn)
1410                 bnxt_re_set_resource_limits(rdev);
1411
1412         rc = bnxt_qplib_alloc_ctx(rdev->en_dev->pdev, &rdev->qplib_ctx, 0,
1413                                   bnxt_qplib_is_chip_gen_p5(&rdev->chip_ctx));
1414         if (rc) {
1415                 pr_err("Failed to allocate QPLIB context: %#x\n", rc);
1416                 goto disable_rcfw;
1417         }
1418         rc = bnxt_re_net_stats_ctx_alloc(rdev,
1419                                          rdev->qplib_ctx.stats.dma_map,
1420                                          &rdev->qplib_ctx.stats.fw_id);
1421         if (rc) {
1422                 pr_err("Failed to allocate stats context: %#x\n", rc);
1423                 goto free_ctx;
1424         }
1425
1426         rc = bnxt_qplib_init_rcfw(&rdev->rcfw, &rdev->qplib_ctx,
1427                                   rdev->is_virtfn);
1428         if (rc) {
1429                 pr_err("Failed to initialize RCFW: %#x\n", rc);
1430                 goto free_sctx;
1431         }
1432         set_bit(BNXT_RE_FLAG_RCFW_CHANNEL_EN, &rdev->flags);
1433
1434         /* Resources based on the 'new' device caps */
1435         rc = bnxt_re_alloc_res(rdev);
1436         if (rc) {
1437                 pr_err("Failed to allocate resources: %#x\n", rc);
1438                 goto fail;
1439         }
1440         set_bit(BNXT_RE_FLAG_RESOURCES_ALLOCATED, &rdev->flags);
1441         rc = bnxt_re_init_res(rdev);
1442         if (rc) {
1443                 pr_err("Failed to initialize resources: %#x\n", rc);
1444                 goto fail;
1445         }
1446
1447         set_bit(BNXT_RE_FLAG_RESOURCES_INITIALIZED, &rdev->flags);
1448
1449         if (!rdev->is_virtfn) {
1450                 rc = bnxt_re_setup_qos(rdev);
1451                 if (rc)
1452                         pr_info("RoCE priority not yet configured\n");
1453
1454                 INIT_DELAYED_WORK(&rdev->worker, bnxt_re_worker);
1455                 set_bit(BNXT_RE_FLAG_QOS_WORK_REG, &rdev->flags);
1456                 schedule_delayed_work(&rdev->worker, msecs_to_jiffies(30000));
1457         }
1458
1459         rtnl_unlock();
1460         locked = false;
1461
1462         /* Register ib dev */
1463         rc = bnxt_re_register_ib(rdev);
1464         if (rc) {
1465                 pr_err("Failed to register with IB: %#x\n", rc);
1466                 goto fail;
1467         }
1468         set_bit(BNXT_RE_FLAG_IBDEV_REGISTERED, &rdev->flags);
1469         dev_info(rdev_to_dev(rdev), "Device registered successfully");
1470         ib_get_eth_speed(&rdev->ibdev, 1, &rdev->active_speed,
1471                          &rdev->active_width);
1472         set_bit(BNXT_RE_FLAG_ISSUE_ROCE_STATS, &rdev->flags);
1473         bnxt_re_dispatch_event(&rdev->ibdev, NULL, 1, IB_EVENT_PORT_ACTIVE);
1474         bnxt_re_dispatch_event(&rdev->ibdev, NULL, 1, IB_EVENT_GID_CHANGE);
1475
1476         return 0;
1477 free_sctx:
1478         bnxt_re_net_stats_ctx_free(rdev, rdev->qplib_ctx.stats.fw_id);
1479 free_ctx:
1480         bnxt_qplib_free_ctx(rdev->en_dev->pdev, &rdev->qplib_ctx);
1481 disable_rcfw:
1482         bnxt_qplib_disable_rcfw_channel(&rdev->rcfw);
1483 free_ring:
1484         type = bnxt_qplib_get_ring_type(&rdev->chip_ctx);
1485         bnxt_re_net_ring_free(rdev, rdev->rcfw.creq_ring_id, type);
1486 free_rcfw:
1487         bnxt_qplib_free_rcfw_channel(&rdev->rcfw);
1488 fail:
1489         if (!locked)
1490                 rtnl_lock();
1491         bnxt_re_ib_unreg(rdev);
1492         rtnl_unlock();
1493
1494         return rc;
1495 }
1496
1497 static void bnxt_re_dev_unreg(struct bnxt_re_dev *rdev)
1498 {
1499         struct bnxt_en_dev *en_dev = rdev->en_dev;
1500         struct net_device *netdev = rdev->netdev;
1501
1502         bnxt_re_dev_remove(rdev);
1503
1504         if (netdev)
1505                 bnxt_re_dev_unprobe(netdev, en_dev);
1506 }
1507
1508 static int bnxt_re_dev_reg(struct bnxt_re_dev **rdev, struct net_device *netdev)
1509 {
1510         struct bnxt_en_dev *en_dev;
1511         int rc = 0;
1512
1513         if (!is_bnxt_re_dev(netdev))
1514                 return -ENODEV;
1515
1516         en_dev = bnxt_re_dev_probe(netdev);
1517         if (IS_ERR(en_dev)) {
1518                 if (en_dev != ERR_PTR(-ENODEV))
1519                         pr_err("%s: Failed to probe\n", ROCE_DRV_MODULE_NAME);
1520                 rc = PTR_ERR(en_dev);
1521                 goto exit;
1522         }
1523         *rdev = bnxt_re_dev_add(netdev, en_dev);
1524         if (!*rdev) {
1525                 rc = -ENOMEM;
1526                 bnxt_re_dev_unprobe(netdev, en_dev);
1527                 goto exit;
1528         }
1529 exit:
1530         return rc;
1531 }
1532
1533 static void bnxt_re_remove_one(struct bnxt_re_dev *rdev)
1534 {
1535         pci_dev_put(rdev->en_dev->pdev);
1536 }
1537
1538 /* Handle all deferred netevents tasks */
1539 static void bnxt_re_task(struct work_struct *work)
1540 {
1541         struct bnxt_re_work *re_work;
1542         struct bnxt_re_dev *rdev;
1543         int rc = 0;
1544
1545         re_work = container_of(work, struct bnxt_re_work, work);
1546         rdev = re_work->rdev;
1547
1548         if (re_work->event != NETDEV_REGISTER &&
1549             !test_bit(BNXT_RE_FLAG_IBDEV_REGISTERED, &rdev->flags))
1550                 return;
1551
1552         switch (re_work->event) {
1553         case NETDEV_REGISTER:
1554                 rc = bnxt_re_ib_reg(rdev);
1555                 if (rc) {
1556                         dev_err(rdev_to_dev(rdev),
1557                                 "Failed to register with IB: %#x", rc);
1558                         bnxt_re_remove_one(rdev);
1559                         bnxt_re_dev_unreg(rdev);
1560                         goto exit;
1561                 }
1562                 break;
1563         case NETDEV_UP:
1564                 bnxt_re_dispatch_event(&rdev->ibdev, NULL, 1,
1565                                        IB_EVENT_PORT_ACTIVE);
1566                 break;
1567         case NETDEV_DOWN:
1568                 bnxt_re_dev_stop(rdev);
1569                 break;
1570         case NETDEV_CHANGE:
1571                 if (!netif_carrier_ok(rdev->netdev))
1572                         bnxt_re_dev_stop(rdev);
1573                 else if (netif_carrier_ok(rdev->netdev))
1574                         bnxt_re_dispatch_event(&rdev->ibdev, NULL, 1,
1575                                                IB_EVENT_PORT_ACTIVE);
1576                 ib_get_eth_speed(&rdev->ibdev, 1, &rdev->active_speed,
1577                                  &rdev->active_width);
1578                 break;
1579         default:
1580                 break;
1581         }
1582         smp_mb__before_atomic();
1583         atomic_dec(&rdev->sched_count);
1584 exit:
1585         kfree(re_work);
1586 }
1587
1588 static void bnxt_re_init_one(struct bnxt_re_dev *rdev)
1589 {
1590         pci_dev_get(rdev->en_dev->pdev);
1591 }
1592
1593 /*
1594  * "Notifier chain callback can be invoked for the same chain from
1595  * different CPUs at the same time".
1596  *
1597  * For cases when the netdev is already present, our call to the
1598  * register_netdevice_notifier() will actually get the rtnl_lock()
1599  * before sending NETDEV_REGISTER and (if up) NETDEV_UP
1600  * events.
1601  *
1602  * But for cases when the netdev is not already present, the notifier
1603  * chain is subjected to be invoked from different CPUs simultaneously.
1604  *
1605  * This is protected by the netdev_mutex.
1606  */
1607 static int bnxt_re_netdev_event(struct notifier_block *notifier,
1608                                 unsigned long event, void *ptr)
1609 {
1610         struct net_device *real_dev, *netdev = netdev_notifier_info_to_dev(ptr);
1611         struct bnxt_re_work *re_work;
1612         struct bnxt_re_dev *rdev;
1613         int rc = 0;
1614         bool sch_work = false;
1615
1616         real_dev = rdma_vlan_dev_real_dev(netdev);
1617         if (!real_dev)
1618                 real_dev = netdev;
1619
1620         rdev = bnxt_re_from_netdev(real_dev);
1621         if (!rdev && event != NETDEV_REGISTER)
1622                 goto exit;
1623         if (real_dev != netdev)
1624                 goto exit;
1625
1626         switch (event) {
1627         case NETDEV_REGISTER:
1628                 if (rdev)
1629                         break;
1630                 rc = bnxt_re_dev_reg(&rdev, real_dev);
1631                 if (rc == -ENODEV)
1632                         break;
1633                 if (rc) {
1634                         pr_err("Failed to register with the device %s: %#x\n",
1635                                real_dev->name, rc);
1636                         break;
1637                 }
1638                 bnxt_re_init_one(rdev);
1639                 sch_work = true;
1640                 break;
1641
1642         case NETDEV_UNREGISTER:
1643                 /* netdev notifier will call NETDEV_UNREGISTER again later since
1644                  * we are still holding the reference to the netdev
1645                  */
1646                 if (atomic_read(&rdev->sched_count) > 0)
1647                         goto exit;
1648                 bnxt_re_ib_unreg(rdev);
1649                 bnxt_re_remove_one(rdev);
1650                 bnxt_re_dev_unreg(rdev);
1651                 break;
1652
1653         default:
1654                 sch_work = true;
1655                 break;
1656         }
1657         if (sch_work) {
1658                 /* Allocate for the deferred task */
1659                 re_work = kzalloc(sizeof(*re_work), GFP_ATOMIC);
1660                 if (re_work) {
1661                         re_work->rdev = rdev;
1662                         re_work->event = event;
1663                         re_work->vlan_dev = (real_dev == netdev ?
1664                                              NULL : netdev);
1665                         INIT_WORK(&re_work->work, bnxt_re_task);
1666                         atomic_inc(&rdev->sched_count);
1667                         queue_work(bnxt_re_wq, &re_work->work);
1668                 }
1669         }
1670
1671 exit:
1672         return NOTIFY_DONE;
1673 }
1674
1675 static struct notifier_block bnxt_re_netdev_notifier = {
1676         .notifier_call = bnxt_re_netdev_event
1677 };
1678
1679 static int __init bnxt_re_mod_init(void)
1680 {
1681         int rc = 0;
1682
1683         pr_info("%s: %s", ROCE_DRV_MODULE_NAME, version);
1684
1685         bnxt_re_wq = create_singlethread_workqueue("bnxt_re");
1686         if (!bnxt_re_wq)
1687                 return -ENOMEM;
1688
1689         INIT_LIST_HEAD(&bnxt_re_dev_list);
1690
1691         rc = register_netdevice_notifier(&bnxt_re_netdev_notifier);
1692         if (rc) {
1693                 pr_err("%s: Cannot register to netdevice_notifier",
1694                        ROCE_DRV_MODULE_NAME);
1695                 goto err_netdev;
1696         }
1697         return 0;
1698
1699 err_netdev:
1700         destroy_workqueue(bnxt_re_wq);
1701
1702         return rc;
1703 }
1704
1705 static void __exit bnxt_re_mod_exit(void)
1706 {
1707         struct bnxt_re_dev *rdev, *next;
1708         LIST_HEAD(to_be_deleted);
1709
1710         mutex_lock(&bnxt_re_dev_lock);
1711         /* Free all adapter allocated resources */
1712         if (!list_empty(&bnxt_re_dev_list))
1713                 list_splice_init(&bnxt_re_dev_list, &to_be_deleted);
1714         mutex_unlock(&bnxt_re_dev_lock);
1715        /*
1716         * Cleanup the devices in reverse order so that the VF device
1717         * cleanup is done before PF cleanup
1718         */
1719         list_for_each_entry_safe_reverse(rdev, next, &to_be_deleted, list) {
1720                 dev_info(rdev_to_dev(rdev), "Unregistering Device");
1721                 /*
1722                  * Flush out any scheduled tasks before destroying the
1723                  * resources
1724                  */
1725                 flush_workqueue(bnxt_re_wq);
1726                 bnxt_re_dev_stop(rdev);
1727                 /* Acquire the rtnl_lock as the L2 resources are freed here */
1728                 rtnl_lock();
1729                 bnxt_re_ib_unreg(rdev);
1730                 rtnl_unlock();
1731                 bnxt_re_remove_one(rdev);
1732                 bnxt_re_dev_unreg(rdev);
1733         }
1734         unregister_netdevice_notifier(&bnxt_re_netdev_notifier);
1735         if (bnxt_re_wq)
1736                 destroy_workqueue(bnxt_re_wq);
1737 }
1738
1739 module_init(bnxt_re_mod_init);
1740 module_exit(bnxt_re_mod_exit);